1/ I’m building a Bittensor SN50 (Synth) miner in public. 🧵
The Synth subnet is a unique challenge in decentralized finance. It doesn't want you to "predict one number." It wants you to produce a distribution of plausible futures.
Here is the roadmap for how I’m doing it:
We’ve been selected for @Salesforce Launchpad.
Big moment for us at Leadpoet. Excited to build alongside this network and accelerate our momentum.
Still early.
@EasyTaoX@markjeffrey@ridges_ai I use for particular issues and it works well - but with the autonomy of other frameworks like claude code, hermes agent etc it creates some friction for the user. Once they build a complete framework it will much better
Here's a visual (estimate) to show what I've been working on and why. Minimum "island" size for distributed training gets out of hand pretty quick, and this is with ternary weight estimates not even fp8/bf16.
Finding 400+ available b200s colocated is hard enough, and that's the floor. DC projects getting cancelled, GPU shortages, power constraints, etc. Sure there are tricks to reduce it, but at the cost of performance.
Instead, what if you could have single 8x b200 nodes per DC (and many DCs), vs hundreds of b200s per island (which all need to be colocated).
Will this algorithm scale perfectly to a trillion parameters or higher? Well, I don't know yet, but it's 100% worth the effort to build it and see. Imagine the possibilities...
Official Minotaur Roadmap Released
https://t.co/wSDRebPXVb
We are proceeding with mainnet testing this week.
The roadmap breaks down each phase with expected timelines and covers what we have built so far. It also reflects the recent shift in scope as we expand toward agentic DeFi and more general intent solving.
We are still focused on the same core problem, just thinking about it at a larger scale. What started as the concept for the subnet is now taking form as our first app powered by Minotaur, a cross-chain DEX aggregator.
Looking forward to seeing this beast come to life.
@Msulabs Its in initial R&D mode right now it take a bit to "warm up" to competitive models. I will post an updates with its progress and will share the experiments so any one can ingest that information and run there own pipeline to find there own models
@NTheEnigma Currently it set up to use lower cost GPUs on basilica (v100, RTXs) - looking at $ 0.05-0.45 per hour. Models are quite small, could you probably batch train a couple models to squeeze most of these GPUs.